WUHAN — Chinese local governments began racing to build humanoid robot centers in October 2023. At the time, China's Ministry of Industry and Information Technology released its first comprehensive road map for nurturing the humanoid robot industry, the "Guiding Opinions on the Innovative Development of Humanoid Robots," which called for strengthening public technology infrastructure. Just 14 months later, 22 innovation centers had opened across the country. That is an unusually fast pace, considering it took two to three years for the first centers to launch after policy announcements in the electric vehicle and semiconductor sectors.
Three years in, concerns about overinvestment remain widespread. Critics say the missteps local governments made in the battery and solar industries are being repeated in robotics. Most local governments operate the centers in partnership with robot makers, buying those companies' robots and producing training data. Operating costs are covered by selling the data externally, but demand is not yet sufficient. The humanoid robot training center in Beijing's Shijingshan district recently terminated its contract with partner company RealMan, citing data quality that fell short of expectations and weak sales revenue.
Analysts say China recognizes the contradiction but is accepting it in the interest of speed. The aim is to secure vast amounts of data first and gain an edge over the United States in the contest for technological supremacy. The center of gravity in the robotics competition between the two countries is shifting from hardware to the robot's "brain." Unlike large language models such as ChatGPT, robot AI models cannot simply draw on internet text and images. They must instead accumulate data from real physical interactions, one piece at a time, in the field. In a report written after a recent visit to a Chinese robotics industry exhibition, Samsung Securities concluded that "the real bottleneck remains in the cerebrum."
Industry officials say that while it is still very early, a "ChatGPT moment" could arrive in robotics once enough data accumulates. Chen Tao, director of the Deep Learning Research Institute at Fudan University, said the biggest problem with current robot vision-language-action, or VLA, models is that they have plenty of success data but too little failure data. "True self-evolution is possible only when a robot corrects its own errors and converts them into experience," Chen said.
China is already ahead in the race to amass data. According to U.S. data-labeling firm Scale AI, China accounts for about 90% of commercially available robot AI data, and its data production costs are 60% lower than in the United States. The sheer number of robots China mass-produces is overwhelming, and it also holds an advantage in the workforce available to train them and in labor costs. Guo Ping, chairman of Huawei's supervisory board, recently cited computing infrastructure, data and talent as the three pillars of AI competition, saying China lags the United States in computing infrastructure but leads on data.
